Snap, Crackle, Pop: Joint Sounds in JIAdcwhittingslow.com/media/presentations/PRoSTalk2019.pdfSnap, Crackle, Pop: Joint Sounds in JIA Investigating Acoustic Emissions from the Joints

Snap, Crackle, Pop: Joint Sounds in JIA

Investigating Acoustic Emissions from the Joints as a Biomarker for JIADaniel Whittingslow, MD/PhD CandidateOmer Inan, PhDDr. Sampath Prahalad, MD

2

JAMS Enrollment – Emory and CHoA

• Children 6-18 years old• 3 Groups:

– Active JIA– JIA Post-Treatment (minimum 6 weeks)– Healthy Controls

Group: # Enrolled: Age (years): Male FemaleJIA 25 12.23 ± 3.1 5 20

JIA Follow-up 12 12.91 ± 2.7 1 11Healthy Controls 18 12.50 ± 3.2 3 15

3

Method

• We attach two contact microphones and an IMU onto each knee.

• The patient flexes/extends their leg 10 times.

• Recorded sounds are analyzed for patterns that could:– Differentiate JIA from HCs– Monitor progression of JIA 2 s

4

Example Sound Recordings

Representative time domain AE signals of four FE repetitions. Healthy controls (HC) have virtually no sounds, JIA patient have repetitive click with a more heterogenous signal, and the follow-up returns toward healthy.

Active JIA JIA Post-TreatmentHealthy Control

Time (s)

Acc

eler

atio

n (m

m/s

2 )

5

Signal Analysis

x 10

Feature Extractionf1 f2 . . . . f49

1

2

3

.

.

.

n

Cyc

les

Num

12...

3637

Y

11...00

LogisticRegression

PJIA

0.80.9

.

.

.0.30.2

Feature Matrix = feature per cycle, Num = subject numbers, Y = ground truth JIA statusPJIA= Probability Estimate of JIA

6

LOSO-CV Accuracy Calculation

Training Set

Testing Set

1 Subject

All Subjects - 1

LogisticRegression

PJIA

0.80.90.70.60.90.30.2

Probability of JIA

𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾 𝐻𝐻𝐾𝐾𝐻𝐻𝐻𝐻𝐻𝐻𝐻 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝐾𝐾 =∑𝑃𝑃𝑆𝑆𝑆𝑆𝑃𝑃𝐻𝐻𝑃𝑃𝑃𝑃𝐻𝐻𝑃𝑃𝐻𝐻𝑃𝑃𝐾𝐾𝑃𝑃

# 𝑆𝑆𝑜𝑜 𝐶𝐶𝐶𝐶𝑆𝑆𝐻𝐻𝐾𝐾𝑃𝑃

𝑱𝑱𝑱𝑱𝑱𝑱 𝒊𝒊𝒊𝒊 𝑲𝑲𝑲𝑲𝑲𝑲𝑲𝑲 𝑯𝑯𝑲𝑲𝑯𝑯𝑯𝑯𝑯𝑯𝑯𝑯 𝑺𝑺𝑺𝑺𝑺𝑺𝑺𝑺𝑲𝑲 > 𝟎𝟎.𝟓𝟓𝐨𝐨𝐨𝐨

𝑯𝑯𝑲𝑲𝑯𝑯𝑯𝑯𝑯𝑯𝑯𝑯𝑯𝑯 𝒊𝒊𝒊𝒊 𝑲𝑲𝑲𝑲𝑲𝑲𝑲𝑲 𝑯𝑯𝑲𝑲𝑯𝑯𝑯𝑯𝑯𝑯𝑯𝑯 𝑺𝑺𝑺𝑺𝑺𝑺𝑺𝑺𝑲𝑲 ≤ 𝟎𝟎.𝟓𝟓

𝐴𝐴𝑆𝑆𝑆𝑆𝐴𝐴𝑆𝑆𝐻𝐻𝑆𝑆𝐶𝐶 =𝐶𝐶𝐶𝐶𝑆𝑆𝐻𝐻𝐾𝐾𝑃𝑃 𝐿𝐿𝐻𝐻𝑃𝑃𝐾𝐾𝐻𝐻𝐾𝐾𝐿𝐿 𝐶𝐶𝑆𝑆𝑆𝑆𝑆𝑆𝐾𝐾𝑆𝑆𝐻𝐻𝐻𝐻𝐶𝐶

𝑇𝑇𝑆𝑆𝐻𝐻𝐻𝐻𝐻𝐻 𝐶𝐶𝐶𝐶𝑆𝑆𝐻𝐻𝐾𝐾𝑃𝑃

Repeat for each subject . . .

7

Cycle Prediction Accuracy Per Subject

Most subjects had >70% cycle labeling accuracy. Overall Accuracy = 81.7%

8

Subject Knee Scores Distribution

9

Model Classification Performance

Specificity = 78.9%Sensitivity = 84.0%

ROC - Area Under Curve = 89.7%)

10

Feature Importance

11

Breaking the Model

12

Tracking the Follow-ups

All but one of the follow-ups showed improvement in their joint health score.The outlier also did not show clinical improvement at 2nd visit.

• Joint sounds show promise for screening, diagnosing, and tracking JIA.

• We should continue recruitment efforts to ensure this model generalizes.

• The feature selection and number of cycles recorded both impact the accuracy of joint sound analysis.

13

Conclusions

14

Thanks, Questions?

Daniel [email protected]